• Joined on 2025-12-11

llama-cpp-vulkan-mtp-batch-checkpoints (pr22673)

Published 2026-05-08 20:33:27 +00:00 by dan

Installation

docker pull gitea.coffee-anon.com/dan/llama-cpp-vulkan-mtp-batch-checkpoints:pr22673
sha256:364f116a24041c647d0265fdbb08ca81824fa61e2b9a407eecd2e985ffd72784

Image Layers

KIWI 10.3.0
RUN /bin/sh -c microdnf -y --nodocs --setopt=install_weak_deps=0 install bash ca-certificates libatomic libstdc++ libgcc vulkan-loader vulkan-loader-devel vulkaninfo mesa-vulkan-drivers radeontop procps-ng && microdnf clean all && rm -rf /var/cache/dnf/* # buildkit
COPY /usr/ /usr/ # buildkit
COPY /usr/local/ /usr/local/ # buildkit
COPY /opt/llama.cpp/build/bin/rpc-* /usr/local/bin/ # buildkit
RUN /bin/sh -c echo "/usr/local/lib" > /etc/ld.so.conf.d/local.conf && echo "/usr/local/lib64" >> /etc/ld.so.conf.d/local.conf && ldconfig && cp -n /usr/local/lib/libllama*.so* /usr/lib64/ 2>/dev/null || true && ldconfig # buildkit
COPY gguf-vram-estimator.py /usr/local/bin/gguf-vram-estimator.py # buildkit
RUN /bin/sh -c chmod +x /usr/local/bin/gguf-vram-estimator.py # buildkit
CMD ["/bin/bash"]
LABEL maintainer=citizendaniel
LABEL description=llama.cpp with MTP (PR #22673) + batch checkpoints (PR #19970) on Vulkan RADV
LABEL patch.pr.mtp=https://github.com/ggml-org/llama.cpp/pull/22673
LABEL patch.pr.batch_checkpoints=https://github.com/ggml-org/llama.cpp/pull/19970
LABEL patch.purpose=A/B bench MTP speculative decoding on Qwen3.6-35B-A3B (hybrid SSM/MoE)
LABEL patch.removal=Revert to a stock build once PR #22673 + #19970 land upstream
COPY /staging/usr/bin/llama-* /usr/bin/ # buildkit
COPY /staging/usr/lib64/libllama* /usr/lib64/ # buildkit
COPY /staging/usr/lib64/libggml* /usr/lib64/ # buildkit
COPY /staging/usr/share/llama.cpp /usr/share/llama.cpp # buildkit
RUN /bin/sh -c ldconfig # buildkit
RUN /bin/sh -c echo "=== MTP + batch-checkpoints overlay verification ===" && ls -la /usr/bin/llama-server && ls -la /usr/lib64/libllama* 2>/dev/null && ls -la /usr/lib64/libggml* 2>/dev/null && echo "=== Binary check ===" && llama-server --version 2>&1 || true # buildkit

Labels

Key Value
description llama.cpp with MTP (PR #22673) + batch checkpoints (PR #19970) on Vulkan RADV
io.buildah.version 1.43.1
license MIT
maintainer citizendaniel
name fedora-minimal
org.opencontainers.image.license MIT
org.opencontainers.image.licenses MIT
org.opencontainers.image.name fedora-minimal
org.opencontainers.image.title fedora-minimal
org.opencontainers.image.url https://fedoraproject.org/
org.opencontainers.image.vendor Fedora Project
org.opencontainers.image.version 43
patch.pr.batch_checkpoints https://github.com/ggml-org/llama.cpp/pull/19970
patch.pr.mtp https://github.com/ggml-org/llama.cpp/pull/22673
patch.purpose A/B bench MTP speculative decoding on Qwen3.6-35B-A3B (hybrid SSM/MoE)
patch.removal Revert to a stock build once PR #22673 + #19970 land upstream
vendor Fedora Project
version 43
Details
Container
2026-05-08 20:33:27 +00:00
14
OCI / Docker
linux/amd64
MIT
634 MiB
Versions (1) View all
pr22673 2026-05-08